Disentangling based Environment-Robust Feature Learning for Person ReID
Yifan Liu (Tsinghua University), Ya-Li Li (Tsinghua University), Shengjin Wang (Tsinghua University)*
The 33rd British Machine Vision Conference
Abstract
Person re-identification (ReID) has received much attention in recent years. As a cross-camera retrieving problem, ReID suffers from the influence of environments. Images captured by the same camera have similar environment features like backgrounds, illuminations and angles, making their features extracted by a neural network have high similarity scores, even if their person IDs are different. A quantitative experiment is designed in this paper to demonstrate the above issue. We proposed a novel Environment-Robust Feature Learning network (EFL) to tackle this problem. First, we designed a feature disentangling module (FDM) based on the idea of minimizing mutual information of identity related features and camera related features. Besides, we adopt a Mutual Mean Teaching (MMT) framework as identity feature extractor to improve the robustness of the features. Moreover, we constructed a multi-environment person ReID dataset ME-ReID (multi-environment) to evaluate our method. Extensive experiments show that our method achieves state-of-the-art performances on widely used datasets Market1501, MSMT17, MARS. Our method also has a great improvement of +9.1\%/+6.9\% of rank1/mAP on ME-ReID, showing the effectiveness of our method. The ME-ReID dataset is available on: https://github.com/liuyf21/ME-ReID-dataset.Video
Citation
@inproceedings{Liu_2022_BMVC,
author = {Yifan Liu and Ya-Li Li and Shengjin Wang},
title = {Disentangling based Environment-Robust Feature Learning for Person ReID},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0428.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by